What is Midjourney’s Style Tuner?
The Style Tuner allows you to train a subset of the Midjourney AI on a specific style with image & text prompts.
This will allow you to generate future images with consistent styles and coloring!
My mind is racing with its potential for creatives!
Possible Use Cases for Designers
Graphic Novel Character Design
Custom Branded photos for a website
Illustrate a children’s book
Quick consistent b-roll assets
- How to Tune a Style in Midjourney?
- How Effective is Style Tuning, Really?
- How to Ensure All Images Follow the Style?
- What Parameters Can you Use with Style Tuning?
- Why Isn’t Tuning Working?
- Can You Share a Style?
How to Tune a Style in Midjourney?
0.) Test Your Prompt
Before you begin tuning a style in Midjourney, you’ll want to test your training prompt.
When you enter your prompt in the tuning command, you actually won’t see the types of pictures the prompt will generate until after it’s used 0.15-1.2 hours of your fast time.
If you test and refine your prompt using the /imagine command then you can ensure your prompt generates the style you are looking for, for significantly less time usage.
IMPORTANT: Style Tuning only works with a small set of parameters, so check that the parameters you use in your prompt are compatible before you move from this step.
1.) Type /tune & Enter Your Prompt
Just as /imagine is the command you use in Midjourney to generate images, /tune is the command used to train styles.
To begin training, type /tune in your Discord chat and press enter to select the command.
Next enter in the prompt you tested in the step before this and press enter again.
2.) Select How Many Style Directions to Train Midjourney With
Next, you’ll be shown a popup that guides you to select how many “Style Directions” you want to use in the training and the style mode you’d like applied to those images.
Midjourney tunes a style by first providing you with a series of images (AKA “Style Directions”) that showcase different elements of your prompt. It then asks you to select images that follow aspects of the style you’d like to key on.
We’ll get into more details surrounding this process later on in the article, but for now, what you need to understand is that you are selecting the number of images you’d like to generate to tune your model. Generating these Style Directions consumes a portion of your fast GPU time, even when you have relaxed mode turned on in the setting. The more images you select, the more time you’ll use from your subscription.
The amount of GPU time /tune uses depends on many factors such as the complexity of your prompt and the amount of directions you’d like to generate, but generally you can expect the following:
GPU Usage Per Style Directions
16 Directions -> 0.15
32 Directions → 0.3
64 Directions → 0.6
128 Directions -> 0.15
On the other side of things, the more images you generate, the more specific you can make the style.
How many style directions you choose to tune your model with depends on what you prioritize more: saving your fast GPU time, or ensuring you tuning a specific style.
As a general rule though, selecting 32 pairs is plenty.
3.) Select Style Parameter (Raw or Default)
Similar to the –style parameter, Midjourney is asking if you’d like to generate images that use Midjourney’s default style or if you’d like to remove stylistic assumptions by using RAW mode.
As a general rule, if you’re generating images you’d like to look like photographs, use RAW mode, otherwise you’re good to use Midjourney’s default mode.
The default mode usually makes the images look more artistic and generally more aesthetic, where as RAW is much more stripped back and therefore more realistic.
4.) Confirm & Generate
Once you’ve selected the number of style directions and your preferred mode, press submit.
This will recalculate how much GPU time your tuning settings will use.
If you are ok with this time being used, click the green button again and Midjourney will start generating your style directions for tuning.
This generation usually takes about 2 minutes.
Once complete, Discord will send you a URL to your training set.
Not seeing the URL? Jump ahead to troubleshooting tips.
5.) Select Training Method
Once your style directions have been generated and you’ve opened your link, there are two different ways you can tune your style:
Compare two styles at a time
Pick your favorite from a big grid
Both methods are similar in the sense you ultimately are selecting images that showcase particular parts of the style you’d like to tune. However, each method allows you a different selection method.
Compare two styles at a time
Using this method, you’ll see a series of image pairs (the same amount of Style Directions you chose), simply select the image that better showcases a quality you like.
You know an image is selected because the white box that was in the middle will have moved to the image you clicked on.
Note: You don’t have to select either image if you don’t like them. You can simply leave those selections with the white box in the middle.
Pick your favorite from a big grid
If you use this method you’ll just see a grid of individual images to select from.
Instead of selecting images based on the style elements Midjourney is comparing, you will just select images that follow the style you are looking to train.
6.) Train Your Style
Once you’ve selected your training method, simply select the images you want to tune your style.
You need to select at least one image in order for a model to be trained.
The more images you select, the more situations the style will cover and the more consistent the style will be.
1 Image Selected
Multiple Images Selected
That being said, only choose the images that will actually help create they style you are looking for. It can be easy to distract the model from what you are looking to create.
7.) Copy Style Code
Once you’ve selected all the images you want to tune your style, scroll down to the bottom of the page and copy the style code that has been generated at the bottom of the page and then open Discord.
Note: If you adjust any image selections, a new code will be generated for you with no GPU cost.
8.) Add Style Code to Future Prompts
Now that you’ve generated your style code, you can now add it to all your new prompts. Simple prompt as you normally would in Discord, but add the –style parameter and paste your code in as the argument.
This should now generate images that use the same style you tuned.
Style not coming through how you imagined for some prompts? This is to be expected when prompts differ largely from your training. Jump ahead to see how you can ensure all photos follow your style.
Optional 9.) Experiment with training
Tuning a model can be a continuous thing! Once you’ve generated your style definitions, you aren’t using any GPU time if you go back and adjust the images that are used to tune the style and generate new codes.
This means there’s nothing stopping you from iterating through a few different style codes and testing which models work best to generate the style you are looking for.
How Effective is Style Tuning Really?
I’ll be blunt, it isn’t perfect yet.
Tuning is really effective when you are generating images that have similar subjects or similar situations.
However, it’s not effective against all prompts. Sometimes your style code doesn’t appear to affect certain images at all.
For example, see how my motel pool and surfer van images mimic the retro style, but the image of the couple sipping drinks doesn’t look similar at all?
This is because both the motel pools and surfer vans can often be associated with the “Motel California” retro aesthetic I was attempting to tune so it’s easy to generate in the same style. However, an image of a couple sipping cocktails is much more general. It’s harder to consistently generate images that don’t have the same direct connection.
Midjourney says it best on their Style Tuner documentation:
“Style codes and prompts interact in complex ways”
It can be hard to predict how styles and prompts will combine since we don’t know what connections the AI is making.
That being said, there are still ways to ensure you are generating images with a similar style every time. Read on to the next section to learn how.
How to Ensure ALL Your Images Follow Your Style
Style Tuning is pretty new, and far from perfect when it comes to adding your style to a new prompt. It doesn’t take you long to find prompts that don’t work with your style code no matter how many style definitions you generated.
The key to ensuring all your images share the same style, is trying to make your new prompts as close as possible to your originals. This way you force Midjourney to see the connection between the images you are trying to generate.
I simply keep my tuning prompt copied in my notes and add keywords back in when I find that a certain prompt isn’t working.
Take my “photo of a couple sipping cocktails –style raw-aP28onvxdBwUskLK” prompt I used in the last section.
Without Extra Keywords
“photo of a couple sipping on cocktails –style raw-aP28onvxdBwUskLK”
“a grainy retro photo of a fashionable couple sipping on cocktails by ludwig favre, cinematic kodortra 400 –style raw-aP28onvxdBwUskLK”
Original Tuning Prompt
If we add keywords from the prompt I used to train the model, I can generate this image in the style I intended.
What Parameters Can You Use with Style Tuning?
As of right now, the /tune command is only compatible with –aspect, –chaos, –tile, and multi prompts.
Why Isn’t Tuning Working?
The first and foremost troubleshooting tip, if the /tune command is not working properly, is to check what version of Midjourney you are using. Style Tuning is only available in version 5.2 or higher.
Is the /tune command not providing you a link? Check that Embeds and Link Previews is turned on in Discords. Simply click on the gear icon next to your name, scroll to App Settings, click on Text & Images, and ensure Embeds & Link Previews is toggled on.